# Before You Trust It: An AI Answer Checklist

<!--
HOW TO USE THIS FILE
- Rename it to something you will recognize (for example: ai-fact-check-checklist.md).
- Keep it next to your desk, or pin it in your team chat.
- Run it whenever an AI answer is about to go somewhere it can do harm:
  an email, a report, a client conversation, anything legal or financial.
- The worked example near the bottom shows one full pass. Replace it with your own.
- This routine lowers your risk. It does not promise to catch everything,
  and that honesty is part of using these tools well.
-->

## Step 0: Does this answer even need a check?

<!-- Skip the routine for brainstorms, rough outlines, and low-stakes questions.
     Run it when a wrong answer would cost you. The quick test is below. -->

The test: if this claim turned out to be wrong after you sent it, would you be
embarrassed or exposed?

- [ ] Yes, run the check (continue below)
- [ ] No, move on

Good candidates for a check:

- [ ] A number, date, or name going into an email, post, or report
- [ ] A claim you plan to repeat to a client, your boss, or an audience
- [ ] A citation, study, or quote the tool offered as evidence
- [ ] Anything legal, financial, medical, or safety-related
- [ ] A summary of a long document you have not read yourself

---

## Step 1: Make it show its sources, then open them

<!-- Do not trust the summary on its own. Ask for the source behind each claim
     you would act on, then click through and confirm the source says it. -->

- [ ] Asked the tool for the source behind each load-bearing claim
- [ ] Opened the links instead of trusting the summary
- [ ] Confirmed each source says what the answer claims it says

## Step 2: Scan for red-flag phrases

<!-- A red flag is not proof of an error, only a prompt to slow down and
     check that one claim. -->

- [ ] "Studies show" with no named study
- [ ] Numbers that are suspiciously round or oddly precise
- [ ] Citations that look real but you cannot find when you search for them

## Step 3: Ground it with search or a second tool

<!-- Re-ask with web search turned on, or paste the same claim into a different
     tool and compare. Agreement raises your confidence. Disagreement shows you
     the spot that needs a closer look. Asking the same model to confirm itself
     often returns the same answer worded differently, so make the second check
     genuinely independent. -->

- [ ] Re-asked with web search on, OR cross-checked in a second, different tool
- [ ] Compared the two answers
- [ ] Two independent answers agree (confidence up) OR disagree (dig in)

## Step 4: Verify the single riskiest claim against a primary source

<!-- Pick the one number, name, date, or quote you would most regret getting
     wrong. Trace it to the original: the filing, the paper, the vendor's own
     documentation, not a blog summarizing it secondhand. -->

- [ ] Picked the one claim I would most regret getting wrong
- [ ] Traced it to the original document, not a secondhand summary
- [ ] Confirmed it word for word

---

## The copy-paste prompt

<!-- Paste this into the same chat after any answer you plan to act on.
     It makes the tool surface its own weak spots instead of defending the answer. -->

```
Before I rely on your last answer, help me check it. Do these four things in
order:

1. List every factual claim you just made (numbers, names, dates, quotes,
   citations) as separate bullet points.
2. For each claim, give the specific source you are basing it on. If it came
   from a web result, include the link.
3. Mark the claims you are least confident about, and say why.
4. Flag any claim you cannot point to a source for, and label it clearly as
   unverified.

Do not rewrite or defend the original answer. I just want the claims, the
sources, and your honest confidence on each one.
```

---

## The one-line rule that anchors all of it

> Verify your riskiest claim against a primary source.

<!-- If you remember nothing else, remember this line. One primary-source check
     on your highest-stakes claim does more for your credibility than a light
     skim of the whole answer. -->

---

## Worked example (replace with your own)

<!-- This is one full pass with the specifics blanked out, so you can see the
     shape without copying anyone's facts. Fill the brackets with your own
     answer and run it. -->

**The answer I got:** "[paste the AI claim you plan to rely on here], and the
tool said a study backs it up."

- Step 1, sources opened: the tool offered no link for the claim. Flag it.
- Step 2, red flags: a study is named in passing with no way to find it, and a
  figure in the answer reads as suspiciously round. Two flags on one sentence.
- Step 3, ground it: re-asked with web search on. The second pass could not
  point me to the named study. Disagreement, so I dig in.
- Step 4, riskiest claim: the figure is the part I would repeat to my boss, so
  it is the riskiest. I went looking for the original document behind it
  instead of trusting the tool's summary.
- What I did with it: kept the parts I could open and confirm, set aside the
  unsourced figure until I find a primary source for it, and labeled the rest
  unconfirmed.

---

<!-- One honest note: even a careful run-through lowers your risk rather than
     removing it. The tool gives you a strong, fast draft. You stay the one who
     decides what is true enough to act on. That is the whole point. -->

*Companion to "How to Fact-Check What AI Tells You" from LaunchReady.ai. No signup. Rename it, share it, run it.*